Speaker Verification Channel Compensation Based on DAE-RBM-PLDA

Abstract

In the speaker recognition system, a model combining the Deep Neural Network (DNN), Identity Vector (I-Vector) and Probabilistic Linear Discriminant Analysis (PLDA) proved to be very effective. In order to further improve the performance of PLDA recognition model, the Denoising Autoencoder (DAE) and Restricted Boltzmann Machine (RBM) and the combination of them (DAE-RBM) are applied to the channel compensation on PLDA model, the aim is to minimize the effect of the speaker i-vector space channel information. The results of our experiment indicate that the Equal Error Rate (EER) and the minimum Detection Cost Function (minDCF) of DAE-PLDA and RBM-PLDA are significantly reduced compared with the standard PLDA system. The DAE-RBM-PLDA which combined the advantages of them enables system identification performance to be further improved.